Enhanced covertness class discriminative universal adversarial perturbations.

Area of Science:

• Computer Vision
• Machine Learning Security
• Adversarial Machine Learning

Background:

• Existing class discriminative universal adversarial perturbations (CD-UAPs) have high fooling ratios for non-targeted classes and are easily detected.
• Current CD-UAP strategies risk discovery due to their impact on non-targeted source classes.

Purpose of the Study:

• To develop a training framework for generating enhanced CD-UAPs with improved covertness.
• To extend CD-UAPs from single-targeted to multi-targeted attacks for greater adversarial precision and reduced detectability.

Main Methods:

• Alternating training of targeted and non-targeted source class sets to update perturbations.
• Introduction of logit pairing to minimize perturbation influence on non-targeted classes.
• Extension of CD-UAPs to multi-targeted attacks, perturbing one source class to multiple sink classes.

Main Results:

• The proposed method generates more deceptive perturbations, enhancing CD-UAP covertness.
• Significant improvements in fooling ratio gaps compared to baseline methods on CIFAR-10, CIFAR-100, and ImageNet datasets.
• Successful demonstration of multi-targeted attacks with high fooling ratios on the GTSRB dataset.

Conclusions:

• The novel training framework effectively enhances the covertness of CD-UAPs.
• The extension to multi-targeted attacks offers precise adversarial control with reduced detection risk.
• This research poses a significant threat to security-sensitive AI applications.